Migrating Alibaba Cloud Chatbot to Cloud Native: Dubbo 3.0 & MSE Governance

Background and Motivation

Alibaba Cloud’s “Cloud‑Xiaomi” dialogue‑robot product, built on deep machine‑learning, NLU, and dialogue‑management technologies, had been released on public cloud since 2017. Rapid growth and the need for stable, frequent releases across public and hybrid clouds exposed inefficiencies in the existing release process.

Challenges

Long (2‑3 weeks) effort each release to analyze module dependencies and simulate upgrade order.

Need for coordinated front‑end, API, and back‑end version alignment to avoid user‑visible errors.

Lack of redundancy and traffic‑governance mechanisms prevented blue‑green or canary deployments.

Migration to Alibaba Cloud

In September 2021 the team migrated the service to Alibaba Cloud, moving from an internal HSF framework to an open‑source stack while maintaining interoperability between on‑premise and cloud environments.

Dubbo 3.0 Adoption

The core requirements were:

Prefer open‑source solutions for easy promotion.

Ensure network‑level security.

Keep upgrade costs low.

The chosen solution upgraded the entire stack to Dubbo 3.0, enabling:

Full‑link upgrade with transparent forwarding; gateway latency < 1 ms.

HTTP/2 support and mTLS between cloud‑native gateways for security.

Multi‑registry discovery across domains, requiring no code changes on the business side.

SDK upgrade to Dubbo 3.0, allowing simple Triple‑service publishing.

Traffic Governance Design

After achieving service registration and discovery, the team focused on traffic governance. The design goals were:

Isolate traffic so that new and old versions run in separate clusters during migration.

Maintain smooth console experience despite front‑end, back‑end, and API version differences.

Allow applications without release requirements to skip the process.

Minimize resource consumption.

Various options were evaluated (internal full‑link solution, unit‑based solution, separate clusters, self‑built isolation). The final approach built a self‑managed isolation layer using RPC routing rules.

Gray‑Release and Blue‑Green Strategies

Two release mechanisms were implemented:

Gray release: Selected applications update individually, using POP’s gray‑traffic routing based on percentage or UID.

Blue‑green release: Deploy a gray cluster for testing, switch all traffic to it, then deploy a base cluster and switch back, finally decommission the gray cluster.

Key configuration steps included setting Linux environment variables (e.g., alicloud.service.tag=gray) and JVM options ( -Dmse.enable=true) to activate MSE traffic‑governance features.

Full‑Link Implementation Results

The first post‑migration release took about 2.5 hours, a significant improvement over previous overnight releases. MSE’s micro‑service governance provided enterprise‑grade full‑link gray capabilities with minimal code changes, leveraging JavaAgent routing.

Future Enhancements

The team plans to extend the solution to cover additional middleware such as RocketMQ and SchedulerX, further enriching the micro‑service governance ecosystem.

Conclusion

By migrating to Alibaba Cloud, upgrading to Dubbo 3.0, and adopting MSE governance, the Cloud‑Xiaomi team achieved faster, safer releases and established a reusable best‑practice for cloud‑native migration and traffic management.